mrcal-convert-lensmodel - Converts a camera model from one lens model to another
$ mrcal-convert-lensmodel
--viz LENSMODEL_OPENCV4 left.cameramodel
... lots of output as the solve runs ...
RMS error of the solution: 3.40256580058 pixels.
Wrote 'left-LENSMODEL_OPENCV4.cameramodel'
... a plot pops up showing the differences ...Given one camera model, this tool computes another camera model that represents the same camera, but utilizes a different lens model. While lens models all exist to solve the same problem, the different representations don't map to one another perfectly, and this tool finds the best-fitting parameters of the target lens model. Two different methods are implemented:
1. If the given cameramodel file contains optimization_inputs, then we have all the data that was used to compute this model in the first place, and we can re-run the original optimization, using the new lens model. This is the default behavior, and is the preferred choice. However it can only work with models that were computed by mrcal originally. We re-run the full original solve, even it contained multiple cameras, unless --monocular is given. With that option, we re-solve only the subset of the images observed by the one requested camera
2. We can sample a grid of points on the imager, unproject them to observation vectors in the camera coordinate system, and then fit a new camera model that reprojects these vectors as closely to the original pixel coordinates as possible. This can be applied to models that didn't come from mrcal. Select this mode by passing --sampled.
Since camera models (lens parameters AND geometry) are computed off real pixel observations, the confidence of the projections varies greatly across the imager and across observation distances. The first method uses the original data, so it implicitly respects these uncertainties: uncertain areas in the original model will be uncertain in the new model as well. The second method, however, doesn't have this information: it doesn't know which parts of the imager and space are reliable, so the results suffer.
As always, the intrinsics have some baked-in geometry information. Both methods optimize intrinsics AND extrinsics, and output cameramodels with updated versions of both. If --sampled: we can request that only the intrinsics be optimized by passing --intrinsics-only.
Also, if --sampled and not --intrinsics-only: we fit the extrinsics off 3D points, not just observation directions. The distance from the camera to the points is set by --distance. This can take a comma-separated list of distances to use. It's STRONGLY recommended to ask for two different distances:
- A "near" distance: where we expect the intrinsics to have the most accuracy. At the range of the chessboards, usually
- A "far" distance: at "infinity". A few km is good usually.
The reason for this is that --sampled solves at a single distance aren't sufficiently constrained. If we ask for a single far distance: "--distance 1000" for instance, we can easily get an extrinsics shift of 100m. This is aphysical: changing the intrinsics could shift the camera origin by a few mm, but not 100m. Conceptually we want to perform a rotation-only extrinsics solve, but this isn't yet implemented. Passing both a near and far distance appears to constrain the extrinsics well in practice. The computed extrinsics transform is printed on the console, with a warning if an aphysical shift was computed. Do pay attention to the console output.
Sampled solves are sometimes sensitive to the optimization seed. To control for this, pass --num-trials to evaluate the solve multiple times from different random seeds, and to pick the best one. These solves are usually quick, so there's no harm in passing something like "--num-trials 10".
We need to consider that the model we're trying to fit may not fit the original model in all parts of the imager. Usually this is a factor when converting wide-angle cameramodels to use a leaner model: a decent fit will be possible at the center, with more and more divergence as we move towards the edges. We handle this with the --where and --radius options to allow the user to select the area of the imager that is used for the fit: observations outside the selected area are thrown out. This region is centered on the point given by --where (or at the center of the imager, if --where is omitted). The radius of this region is given by --radius. If '--radius 0' then we use ALL the data. A radius<0 can be used to set the size of the no-data margin at the corners; in this case I'll use
r = sqrt(width^2 + height^2)/2. - abs(radius)There's a balance to strike here. A larger radius means that we'll try to fit as well as we can in a larger area. This might mean that we won't fit well anywhere, but we won't do terribly anywhere, either. A smaller area means that we give up on the outer regions entirely (resulting in very poor fits there), but we'll be able to fit much better in the areas that remain. Generally empirical testing is required to find a good compromise: pass --viz to see the resulting differences. Note that --radius and --where applies only if we're optimizing sampled reprojections; if we're using the original optimization inputs, the options are illegal.
The output is written to a file on disk, with the same filename as the input model, but with the new lensmodel added to the filename.
to The target lens model
model Input camera model. If omitted or "-", we read
standard input and write to standard output-h, --help show this help message and exit
--sampled Instead of solving the original calibration problem
using the new lens model, use sampled imager points.
This produces biased results, but can be used even if
the original optimization_inputs aren't available
--gridn GRIDN GRIDN Used if --sampled. How densely we should sample the
imager. By default we use a 30x20 grid
--distance DISTANCE Required if --sampled and not --intrinsics-only. A
sampled solve fits the intrinsics and extrinsics to
match up reprojections of a grid of observed pixels.
The points being projected are set a particular
distance (set by this argument) from the camera. Set
this to the distance that is expected to be most
confident for the given cameramodel. Points at
infinity aren't supported yet: specify a high distance
instead. We can fit multiple distances at the same
time by passing them here in a comma-separated,
whitespace-less list. If multiple distances are given,
we fit the model using ALL the distances, but --viz
displays the difference for the FIRST distance given.
See the description above. Without --sampled, this is
used for the visualization only
--intrinsics-only Used if --sampled. By default I optimize the
intrinsics and extrinsics to find the closest
reprojection. If for whatever reason we know that the
camera coordinate system was already right, or we need
to keep the original extrinsics, pass --intrinsics-
only. The resulting extrinsics will be the same, but
the fit will not be as good. In many cases, optimizing
extrinsics is required to get a usable fit, so
--intrinsics-only may not be an option if accurate
results are required.
--where WHERE WHERE Used with or without --sampled. I use a subset of the
imager to compute the fit. The active region is a
circle centered on this point. If omitted, we will
focus on the center of the imager
--radius RADIUS Used with or without --sampled. I use a subset of the
imager to compute the fit. The active region is a
circle with a radius given by this parameter. If
radius == 0, I'll use the whole imager for the fit. If
radius < 0, this parameter specifies the width of the
region at the corners that I should ignore: I will use
sqrt(width^2 + height^2)/2. - abs(radius). This is
valid ONLY if we're focusing at the center of the
imager. By default I ignore a large-ish chunk area at
the corners.
--monocular Used if not --sampled. By default we re-solve the full
calibration problem that was used to originally obtain
the input model, even if it contained multiple
cameras. If --monocular, we re-solve only a subset of
the original problem, using ONLY the observations made
by THIS camera
--viz Visualize the differences between the input and output
models. If we have --distance, the FIRST given
distance is used to display the fit error
--cbmax CBMAX Maximum range of the colorbar
--title TITLE Used if --viz. Title string for the diff plot.
Overrides the default title. Exclusive with
--extratitle
--extratitle EXTRATITLE
Used if --viz. Additional string for the plot to
append to the default title. Exclusive with --title
--hardcopy HARDCOPY Used if --viz. Write the diff output to disk, instead
of making an interactive plot
--terminal TERMINAL Used if --viz. gnuplotlib terminal. The default is
good almost always, so most people don't need this
option
--set SET Used if --viz. Extra 'set' directives to gnuplotlib.
Can be given multiple times
--unset UNSET Used if --viz. Extra 'unset' directives to gnuplotlib.
Can be given multiple times
--force, -f By default existing models on disk are not
overwritten. Pass --force to overwrite them without
complaint
--outdir OUTDIR Directory to write the output into. If omitted, we use
the directory of the input model
--num-trials NUM_TRIALS
If given, run the solve more than once. Useful in case
random initialization produces noticeably different
results. By default we run just one trial, which is
enough most of the timehttps://www.github.com/dkogan/mrcal
Dima Kogan, <dima@secretsauce.net>
Copyright (c) 2017-2023 California Institute of Technology ("Caltech"). U.S. Government sponsorship acknowledged. All rights reserved.
Licensed under the Apache License, Version 2.0 (the "License"); You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0